Agency By Default

Agency by default is the posture where the product acts on user intent immediately, shows exactly what it did, and offers a clear undo path instead of nagging with confirmation dialogs at every step. The model stops negotiating and starts executing because the user already expressed the goal through a prompt or command. This turns the AI from a hesitant helper into the actual surface the user reaches for first. In AI native products the model is the load bearing wall so the default behavior must be decisive action paired with immediate visibility. Cursor agent mode Granola note augmentation and Arc Search browse for me all operate this way. They understand that every would you like me to click burns user attention and broadcasts that the team does not trust its own model. The principle demands transparency surfaces right alongside it. The user must see the context the model pulled from the exact changes it made and the sources it referenced or the whole thing collapses into a hostile black box. Without that visibility agency stops feeling empowering and starts feeling like a rogue intern rearranging your desk while you are away. This posture is non negotiable for any product that wants to pass the deletion test. Remove the model and the shell must feel broken because the primary workflow lived inside the agent's actions. Linear ships a quieter version inside its command bar where natural language commands trigger issue creation triage and status updates without extra approval layers. The timeline simply logs the AI action like any other event. The entire approach rewires the user's mental model from learning rigid interfaces to stating intent and reviewing outcomes. Products that internalize this in 2026 will feel alive. Everything else will feel like antique software with extra sparkle icons.

This is not the reckless free for all where the product makes permanent destructive moves with no audit trail or recovery option. It is not the 2024 enterprise pattern of bolting a chat sidebar onto an existing dashboard then forcing the user through three confirmation modals before the model does anything useful. That pattern is permission by default and it kills momentum while screaming distrust in the model. The cautionary tales from that year are littered with CRMs and analytics tools that responded to every request with would you like me to summarize this after the user had already asked for exactly that. Users closed the panel after one session and never returned. Agency by default is also not hiding the work or burying the undo three menus deep. If the diff is not obvious or the source links require detective work then you have not shipped agency you have shipped anxiety wrapped in marketing copy. Many teams tried to add this principle in sprint twelve after the core product was already locked in and they failed every other principle as a result. The article nails it. If the team does not trust the model the user never will and the product ends up with an autonomy slider nobody touches because the friction feels protective at first then cowardly. True agency by default requires the recovery surface to be designed at the exact same time as the action layer. Skip that and you teach users to fear your product instead of leaning on it.

Cursor delivers the cleanest concrete example of agency by default in shipping software. Open a codebase highlight a module and type refactor this for concurrency using the Cmd K prompt. The agent does not pop modals asking permission to touch each file. It reads the surrounding context plans the changes edits multiple files updates related tests and drops a reviewable diff view with line by line explanations. You accept reject or iterate without ever leaving the flow. The transparency surface is the diff itself plus the logged reasoning in the side chat that you can always expand. Granola takes the same posture into meetings. You take manual notes your own way during the call. Afterward the model transcribes the audio augments your bullets with missing context pulls out action items and suggests priorities all without a single permission request. The interface shows raw transcript on the left and augmented version on the right so every addition is visible and one click removable. Arc Search applies it to the entire browsing experience. Tap browse for me on a query and it spins up parallel visits across multiple sites extracts key information synthesizes a single coherent page and anchors source links throughout the content. No step by step approvals no list of blue links just the synthesized result ready for immediate use. Linear uses a quieter version inside its command palette. Type triage these bugs by severity and impact and it parses the thread creates new issues sets priorities assigns owners based on your history and logs the entire action in the timeline like any other update. Each product passes the deletion test on its core surfaces because stripping the model leaves a hollow stub. They succeed because they rebuilt the workflow around the model instead of tacking cautious suggestions onto forms that already worked without it.

Apply agency by default on surfaces where the cost of a mistake is low the value of momentum is high and you have already shipped the transparency and recovery mechanisms. It belongs in code editors note takers search tools report generators and any creative iteration loop where review and tweak is the expected rhythm. Perplexity uses a lighter version by defaulting to full answers with inline citations and follow up questions instead of asking whether you want research performed. Granola leverages it for post meeting processing where real time latency is not a factor so the model can deliver higher quality augmentation before showing its work. Pair it with streaming UI and skeleton states so the first token appears instantly and the user never wonders whether the agent crashed. The pre ship checklist from the article forces you to count confirmation modals on the primary flow. More than one means you are too cautious. Force the model to hallucinate a wrong answer and test the recovery path. If that path is missing or hidden you are not ready to ship. Avoid agency by default for actions that move money delete data send external messages or trigger compliance requirements. In those cases default to explicit permission with clear previews and final consent. The article references the autonomy slider from agent UI design patterns as the control that lets teams tune this over time. Start conservative if your retrieval or model quality is still shaky then raise it as confidence grows. Never treat agency as a feature you sprinkle on later. It must be architected from day one alongside prompt as input and transparency surfaces or the product will feel like 2024 enterprise bloatware with extra buttons nobody clicks. Teams that get this right stop building tools users learn and start building tools users talk to.

Ship the undo surface before you ship the agency or expect users to treat your product like a liar that overpromises.